Improved face mask detection with super-resolution techniques

Abstract

Super-Resolution is the process of reconstructing a low resolution image into a high resolution image. In recent years, many deep learning based techniques have surfaced and as a result, super-resolution has become a competitive field spurring the proposal of many state-of-the-art models. Super-Resolution can potentially have many applications and one such application, which is especially relevant during this COVID-19 pandemic, is face mask detection. Face mask detection has been implemented rapidly around the world since the start of the pandemic and this project shows that super-resolution techniques help improve the accuracy of face mask detection. Three models which are SSD based models enhanced with the addition super-resolution layers are pitted against the baseline model without super-resolution layers present. All models were trained, validated and tested on a dataset containing 14,016 images of masked and unmasked faces. All of the proposed models beat the baseline model’s mean average precision (mAP) of 76.73% where the best mAP achieved was 80.69%.